A healthcare voice AI company · Medical
Checking a voice agent, and the AI judge that scores it
Human evaluation of a voice agent that books patient appointments, including the LLM judge scoring its calls. Client anonymised at their request.
The challenge
A Y Combinator backed company runs voice agents that answer patient calls 24/7, booking appointments and taking intake details through EHR and telephony integrations.
An LLM judge scored every call automatically. The company needed people to check both the agent’s actions and whether the judge itself was right.
What we did
Reviewers trained in medical terminology scored each call on 10 dimensions: booking details, intake accuracy, tool call choice, patient and provider identity, verbal confirmation, clinical information, the judge’s score and the judge’s error type, and overall quality.
For every failure they tagged the first step where it went wrong, classified it as an agent or judge failure, and escalated disagreements with the judge to the client’s ML team.
Results
A structured, auditable human layer above both the agent and its automated judge, with HIPAA relevant errors tracked separately.
